[Kernel] fp8 conv3d: 8-wave GEMM pipeline + BIG_IN fix#860

Draft

jiacao-amd wants to merge 23 commits into

jiacao-amd:jiacao/conv3d-gemm-port

Commits on Jul 8, 2026

conv3d: parameterize tile size + add autotuner
jiacao-amd
committed

Commits on Jul 13, 2026

conv3d: unify 1D/2D/3D behind one entry + merge split-K helpers

jiacao-amd
and
claude
committed

Commits on Jul 14, 2026

conv3d: async global->LDS copy + 4-stage pipeline (+10-14%)
jiacao-amd
committed
conv3d: extend async copy to large (>2^31 elem) tensors
jiacao-amd
committed
conv3d: add 3x1x1 unfold+hgemm fast path
jiacao-amd
committed
conv3d: revert 3x1x1 unfold+hgemm (large-M A materialization causes -76% regression)
jiacao-amd
committed
conv3d: restore 1x1x1 fast path with accurate comment
jiacao-amd
committed
conv3d: rename to conv3d_implicit, force async path, drop sync dead code
jiacao-amd
committed
conv3d: force async path, drop sync dead code, drop async naming
jiacao-amd
committed

Commits on Jul 15, 2026

fp8 conv: add 8-wave GEMM pipeline kernel + fix BIG_IN rebase

jiacao-amd
and
claude
committed

Commits on Jul 16, 2026

conv3d fp8: add 4-wave GEMM pipeline kernel (slower than 8-wave)

jiacao-amd
and
claude
committed
conv3d fp8: fix split-K heuristic + add WGM L2-swizzle to gemm8w
jiacao-amd
committed
conv3d fp8: route aligned shapes to gemm8w fast path + WGM autotune

jiacao-amd
and
claude
committed
conv3d fp8: generalize gemm8w to all shapes (N/K-partial, tiny-K)
jiacao-amd
committed

Commits on Jul 17, 2026

conv3d fp8: replace general kernel with the 8-wave GEMM-pipeline kernel

jiacao-amd
and
claude
committed
conv3d fp8: drop the 4-wave GEMM-pipeline kernel

jiacao-amd
and
claude
committed
conv3d fp8: drop the 4w-vs-8w bench script

jiacao-amd
and
claude
committed
conv3d fp8: fix >2 GB (BIG_IN / BIG_OUT) NaN + memory fault

jiacao-amd
and
claude
committed